Method and Apparatus for Classifying Document Based on Attention Mechanism and Semantic Analysis

1. A method of operating a system for classifying a document, the method comprising: obtaining a plurality of word embeddings from a plurality of words constituting a plurality of sentences included in the document; providing, to a semantic analysis model, the plurality of word embeddings, wherein the semantic analysis model generates, based on the plurality of word embeddings, a plurality of document features representing the document, the plurality of document features including a keyword similarity and a sentence similarity; extracting, from the semantic analysis model, the plurality of document features; providing, to an inference model, the plurality of word embeddings and the plurality of document features, wherein the inference model evaluates the document based on the plurality of word embeddings and the plurality of document features and generates an evaluation result of the document; extracting, from the inference model, the evaluation result; and outputting the evaluation result, wherein, for generating the evaluation result of the document, the inference model performs: generating, using a hierarchical attention network (HAN), a document vector from the plurality of word embeddings; concatenating the keyword similarity and sentence similarity with the document vector to generate a concatenated vector; and generating, using a fully connected layer, the evaluation result of the document based on the concatenated vector.

2. The method of claim 1, further comprising generating the plurality of word embeddings by encoding, using an encoder, the plurality of words, the encoder implemented with a deep learning engine configured to perform unsupervised learning based on an unlabeled large-scale corpus.

3. The method of claim 1, wherein the evaluation result of the document includes an evaluation decision, an evaluation score, or an evaluation grade of the document.

4. The method of claim 1, wherein: for generating the keyword similarity, the semantic analysis model counts a frequency of the plurality of words having a similarity between a word embedding of each word and a preset target keyword being greater than a preset threshold similarity; and for generating the sentence similarity of the plurality of sentences included in the document, the semantic analysis model averages the plurality of word embeddings to generate sentence vectors.

5. The method of claim 1, further comprising: gathering, from the inference model and the semantic analysis model, support data related to the evaluation result and the plurality of document features; and outputting the support data.

6. The method of claim 5, wherein the support data includes at least one of: an important word selected by an order of weights for the plurality of words provided by the HAN; an important sentence selected by an order of weights for the plurality of sentences provided by the HAN; the keyword similarity; and the sentence similarity.

7. The method of claim 1, wherein the inference model is implemented with a deep learning engine configured to perform supervised learning based on a plurality of training documents and a plurality of labels corresponding to the plurality of training documents.

8. A non-transitory computer-readable medium comprising instructions that, when executed by a processor, cause the processor to control a system for classifying a document to perform the method of claim 1.

9. A method for training an inference model for classifying a document, the method comprising: obtaining a plurality of word embeddings from a plurality of words constituting a plurality of sentences included in the document; obtaining a label indicating an evaluation decision, an evaluation score, or an evaluation grade of the document; obtaining, from a semantic analysis model, a plurality of document features representing the document, the plurality of document features including a keyword similarity and a sentence similarity; and providing, to the inference model, the plurality of word embeddings and the plurality of document features, the inference model is configured to perform: generating, based on the plurality of document features and the plurality of word embeddings, an evaluation result of the document, the evaluation result indicating a probability of the evaluation decision, an estimate for the evaluation score, or an estimate for the evaluation grade; defining a loss function based on the evaluation result and the label; and updating a set of parameters of the inference model to decrease the loss function, wherein, for generating the evaluation result, the inference model is configured to perform: generating, using a hierarchical attention network (HAN), a document vector from the plurality of word embeddings; concatenating the keyword similarity and sentence similarity with the document vector to generate a concatenated vector; and generating, using a fully connected layer, the evaluation result of the document based on the concatenated vector.

10. The method of claim 9, further comprising generating the plurality of word embeddings by encoding, using an encoder, the plurality of words, the encoder implemented with a deep learning engine configured to perform unsupervised learning based on an unlabeled large-scale corpus.

11. The method of claim 9, wherein: for generating the keyword similarity, the semantic analysis model counts a frequency of the plurality of words having a similarity between a word embedding of each word and a preset target keyword being greater than a preset threshold similarity; and for generating the sentence similarity of the plurality of sentences included in the document, the semantic analysis model averages the plurality of word embeddings to generate sentence vectors.

12. A non-transitory computer-readable medium comprising instructions that, when executed by a processor, cause the processor to control a system for training an inference model for classifying a document to perform the method of claim 9.

13. A system for classifying a document, comprising: a processor; and a computer-readable medium in communication with the processor and storing instructions that, when executed by the processor, cause the processor to control the system to perform: obtaining a plurality of word embeddings from a plurality of words constituting a plurality of sentences included in the document; providing, to a semantic analysis model, the plurality of word embeddings, wherein the semantic analysis model generates, based on the plurality of word embeddings, a plurality of document features representing the document, the plurality of document features including a keyword similarity and a sentence similarity; extracting, from the semantic analysis model, the plurality of document features; providing, to an inference model, the plurality of word embeddings and the plurality of document features, wherein the inference model evaluates the document based on the plurality of word embeddings and the plurality of document features and determines an evaluation result of the document; extracting, from the inference model, the evaluation result; and outputting the evaluation result, wherein, for generating the evaluation result of the document, the inference model performs: generating, using a hierarchical attention network (HAN), a document vector from the plurality of word embeddings; concatenating the keyword similarity and sentence similarity with the document vector to generate a concatenated vector; and generating, using a fully connected layer, the evaluation result of the document based on the concatenated vector.

14. The system of claim 13, wherein the instructions, when executed by the processor, further cause the processor to control the system to perform encoding, using an encoder, the plurality of words to generate the plurality of word embeddings, the encoder implemented with a deep learning engine configured to perform unsupervised learning based on an unlabeled large-scale corpus.

15. The system of claim 13, wherein the evaluation result includes an evaluation decision, an evaluation score, or an evaluation grade of the document.

16. The system of claim 13, wherein: for generating the keyword similarity, the semantic analysis model counts a frequency of the plurality of words having a similarity between a word embedding of each word and a preset target keyword being greater than a preset threshold similarity; and for generating the sentence similarity, the semantic analysis model averages the plurality of word embeddings to generate sentence vectors.

17. The system of claim 13, wherein the instructions, when executed by the processor, further cause the processor to control the system to perform: gathering, from the inference model and the semantic analysis model, support data related to the evaluation result and the plurality of document features; and outputting the support data.

18. The system of claim 17, wherein the support data includes at least one of: an important word selected by an order of weights for the plurality of words provided by the HAN; an important sentence selected by an order of weights for the plurality of sentences provided by the HAN; the keyword similarity; and the sentence similarity.

19. The system of claim 13, wherein the inference model is implemented with a deep learning engine configured to perform supervised learning based on a plurality of training documents and a plurality of labels corresponding to the plurality of training documents.